metadata
tags:
- regression
- pytorch
license: mit
Model Description
NumAdd-v2.0 is an optimized feed-forward neural network (FNN) in PyTorch for numerical sum prediction.
Architecture: 2-input, 1-output, with two hidden layers (32, 64 neurons) and ReLU activations.
Parameters: 2,273 trainable.
Precision: Requires torch.float64 (double precision).
Training Config: Optimal batch size: 2048, Final tuning learning rate: 1.0e-12.
Evaluation
Benchmarked on 120,000 samples across six input magnitude ranges. Metrics: MAE, MSE, RMSE, R2.
| Range (Input Max) | MAE | MSE | RMSE | R2 |
|---|---|---|---|---|
| 0-50 | 0.004 | 0.000 | 0.004 | 1.000 |
| 51-500 | 0.003 | 0.000 | 0.004 | 1.000 |
| 501-5000 | 0.004 | 0.000 | 0.004 | 1.000 |
| 5001-50000 | 0.004 | 0.000 | 0.005 | 1.000 |
| 50001-500000 | 0.010 | 0.001 | 0.028 | 1.000 |
| 500001-50000000 | 0.706 | 6.333 | 2.517 | 1.000 |
Limitations
Precision degrades for extremely large magnitude inputs (e.g., >500,000), indicated by increased MAE/MSE, although R2 remains high.